Spewer routes cheap coding tasks to local models

Spewer, released by Chetan Conikee on August 29 under an Apache-2.0 license, is a Homebrew-installed CLI that doesn't replace your coding assistant. It sits next to Codex or Claude Code and watches for jobs you could hand to something cheaper: Luna, Qwen3, or another open-weight model running through Ollama.

That matters because every dollar spent on a frontier model for boilerplate work is a dollar not spent on the parts of a codebase that actually need a frontier model's judgment. Whether the routing is worth the setup depends on how much of a given day's work is well-specified enough to delegate.

Spewer's approach comes down to four rules that hold regardless of which frontier tool is driving. Stay inside your existing harness instead of switching tools mid-task. Only delegate work with a defined success criterion, so a script either passes tests or it doesn't. Compare the same type of work across models instead of assuming one is uniformly weaker. Have the expensive model review the cheap one's output before you trust it.

That last rule is what makes the other three tolerable. A local model producing wrong code is fine if something catches it before it ships, and Spewer puts the frontier model in that checking role. It costs tokens, but fewer than writing the code itself would have.

What Spewer doesn't do is tell you how much you'll save. Conikee says he hasn't run it enough times to back a percentage, and says so directly instead of publishing a placeholder number. Most tools in this space lead with a savings claim on the landing page instead.

What's untested: how the review step holds up on larger, less well-specified tasks, where a defined success criterion gets harder to write in the first place.